DeepSeek V5 release remains unconfirmed with 2027 target

Current status of DeepSeek V5

DeepSeek has not officially released, previewed, or announced the V5 model. The company's official news page and API changelog do not contain any entries for a V5 version. The most recent update from the company is the DeepSeek-V4.1-Flash-1017, which was shipped on 17 October 2026.

While many rumors circulate online, the company's own documentation remains the primary source of truth. DeepSeek's Hugging Face account, where it shares open weights, also lacks any repository for a V5 model. Users currently looking for the latest technology should look to the V4.1-Flash model instead.

The company has even warned its community to rely only on its official accounts for news. Statements from other channels do not reflect the company's actual views or plans. This is important because many third-party reports have stretched minor details into significant claims about an imminent launch.

The roadmap to V4.1-Pro

DeepSeek has indicated that the next model in its public lineup is V4.1-Pro. When the company launched V4.1-Flash on 10 September 2026, it described the new model as the smallest in a new architecture family. The company noted that V4.1-Flash was designed to scale to larger models in the future.

The company originally planned to route all V4-Pro traffic to V4.1-Flash until the larger V4.1-Pro sibling arrived. However, DeepSeek later decided to continue providing API services for the older V4-Pro model in response to user demand. This means V4.1-Pro is the immediate next step in the development cycle rather than V5.

The naming of these models matters for developers and researchers who are tracking the company's progress. A V5 model would represent a completely new generation of technology. Because V4.1-Pro is still in development, a V5 release is naturally pushed further down the timeline. This sequence aligns with the company's history of releasing new generations more than a year apart.

Timeline and hardware reports

The most specific information regarding V5 comes from the Chinese financial outlet Caiwen. A report from 3 November 2026 cited industry sources stating that V5 training is still in the selection phase. These sources suggest that a release is aimed for sometime in 2027, depending on how the training process goes.

The report also suggests a significant shift in hardware for the V5 model. It claims that DeepSeek plans to use Huawei Ascend 950 chips as the core training compute instead of Nvidia GPUs. While neither DeepSeek nor Huawei has confirmed this, it would be a major test of whether a top Chinese model can be pretrained on Chinese hardware.

Pretraining is the long and expensive stage where a model learns from raw data. Because this stage takes months, a project that is still in the planning phase cannot ship in the near future. Even a smooth training run starting late in 2026 would likely push a public release well into 2027. This timeline contradicts earlier analyst forecasts from late 2026.

Decoding the 'leaks'

A viral leak recently claimed that V5 names were found in DeepSeek's own code. This refers to the DeepSeek Harness, which is the company's open-source coding agent. However, a technical review shows that these names are merely test placeholders used by the developers.

The code uses 'deepseek-v5' as a fake model name to test how the application handles unknown models. For example, the tests check if the app correctly refuses a model that is not supported or asks the user to upgrade. These tests do not provide any real specifications about the actual V5 model's capabilities.

One of the test files even sets a fake context window of 9 tokens for the V5 name. Since no real model would have such a tiny context window, this confirms the names are just placeholders. Developers built the Harness to accept new models automatically, which will make a future launch smoother, but it confirms nothing about a current release.

What to use today

For users who want the best current performance from DeepSeek, the V4.1-Flash-1017 model is the recommended choice. DeepSeek states that this model beats the older V4-Pro in terms of performance, cost, and speed. It is currently available for free in the DeepSeek web chat and mobile app.

For those using the API, the V4.1-Flash model offers competitive pricing. It costs $0.15 per million input tokens and $0.60 per million output tokens during off-peak hours. These rates are significantly lower than the costs associated with the older V4-Pro model, making it a cost-effective option for developers.

DeepSeek also continues to provide open weights for its models under the MIT license. Users can find the V4.1-Flash weights on Hugging Face to run them in their own environments. While waiting for V5, these tools provide the most up-to-date capabilities currently offered by the company.